A Termite Nest Detection Method for Earth-Rock Dams Based on MSCCEAUNet and Physical Constraints

Abstract

A termite nest detection method for earth-rock dams, driven by MSCCEAUNet and physical constraints, includes the following steps: Simulating a complex geological model of an earth-rock dam structure with termite nests, and generating corresponding b-scan data and underground dielectric constant data as simulation datasets using the finite difference method (FDTD); constructing an MSCCEAUNet network with multi-scale concatenated convolution (MSC) and efficient channel attention (ECA) mechanism, and inputting the complex geological model data generated in step one into the MSCCEAUNet network for training; adding Gaussian noise with different signal-to-noise ratios and random medium disturbance data to train MSCCEAUNet again, and optimizing the MSCCEAUNet structure based on the model's training performance to improve the model's anti-interference ability and robustness; integrating the physical residual equations that satisfy the physical information constraints of Maxwell's equations for ground-penetrating radar as the physical loss function into the MSCCEAUNet framework to achieve joint data-driven inversion.

Description

Technical Field

[0001] This invention relates to the field of geophysical technology, and in particular to intelligent detection technology for termite nests in earth-rock dams. Specifically, it relates to an intelligent detection method for termite nests in earth-rock dams based on GPR intelligent inversion jointly driven by MSCCEAU-Net and physical information constraints. Background Technology

[0002] Ground-penetrating radar (GPR) is a non-destructive detection technology widely used for the detection and imaging of underground targets. It acquires physical information about the interior of the underground medium by observing the propagation and reflection of electromagnetic waves. However, due to the complexity of the underground environment, traditional physical model-based GPR inversion methods may face limitations in accuracy and resolution. Furthermore, GPR inversion often involves issues such as multiple solutions and noise, making it difficult for the inversion results to accurately reflect the actual underground structure.

[0003] These challenges are even more pronounced for termite nest detection within earth-rock dams. Termite nests are often small, complex in shape, and unevenly distributed, and their electromagnetic signatures differ only slightly from the surrounding medium, making them difficult to reliably identify using traditional methods. Furthermore, the spread of termite nests within the dam structure weakens its stability, posing a long-term safety hazard. Therefore, high-precision detection methods for termite nests are crucial for engineering safety monitoring.

[0004] Deep learning methods have demonstrated significant advantages in handling complex nonlinear problems, automatic feature extraction, and efficient computation. Applying deep learning to the GPR inversion problem can not only improve inversion accuracy but also reduce reliance on manual intervention, enhancing the automation and intelligence of data processing. Research in this direction will help drive the intelligent upgrade of GPR technology, providing new solutions for more accurate, rapid, and efficient underground target detection and physical property assessment.

[0005] In deep learning-based ground-penetrating radar (GPR) inversion research, incorporating physical constraints not only improves the accuracy and reliability of the model but also imbues the research with deeper scientific significance. Specifically, combining deep learning with physical laws provides an efficient and accurate solution for the application of GPR in termite nest detection in earth-rock dams. This research is not only of great value in engineering applications but also lays a scientific foundation for the development of intelligent inversion technology.

[0006] Patent document CN117371330A discloses a two-dimensional magnetotelluric inversion method using deep learning. It constructs a magnetotelluric dataset using traditional inversion methods and then uses a deep learning model to perform two-dimensional magnetotelluric inversion. While this method considers the automatic feature extraction and efficient computation capabilities of deep learning inversion, it fails to consider the physical meaning of electromagnetic wave propagation in the medium. This results in the inversion results lacking practical physical meaning in the absence of physical information constraints. Summary of the Invention

[0007] The purpose of this invention is to solve the technical problems of existing detection technologies for termite nests in earth-rock dams, such as high computational complexity, sensitivity to real-world noise, and lack of practical physical meaning in the inversion results due to the lack of physical information constraints in existing inversion techniques.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0009] A termite nest detection method for earth-rock dams, jointly driven by MSCCEAUNet and physical constraints, includes the following steps:

[0010] Step 1: Simulate a complex geological model of an earth-rock dam structure with termite nests, and use the finite difference method (FDTD) to generate corresponding b-scan data and underground dielectric constant data as simulation datasets;

[0011] Step 2: Construct an MSCCEAUNet network with multi-scale concatenated convolution (MSC) and efficient channel attention mechanism (ECA), and input the complex geological model data generated by the simulation in Step 1 into the MSCCEAUNet network for training;

[0012] Step 3: Add Gaussian noise and random medium perturbation data with different signal-to-noise ratios to train MSCCEAUNet again. Based on the training results, optimize the MSCCEAUNet structure to improve the model's anti-interference ability and robustness.

[0013] Step 4: Integrate the physical residual equations that satisfy the physical information constraints of Maxwell's equations for ground-penetrating radar into the MSCCEAUNet framework as physical loss functions to achieve joint data-driven inversion.

[0014] Step 5: Add data with different signal-to-noise ratios of Gaussian noise and random medium perturbations again to perform joint data and physics-driven training, and determine the optimal weights for each part of the data loss, physical loss and structural similarity loss of MSCCEAUNet.

[0015] In step one, the gprMax simulation software is used to simulate the earth-rock dam structure model with termite nests.

[0016] In step four, when jointly driving the model and physical information constraints, the following steps are adopted:

[0017] Step 4-1) First, decouple the electric field and magnetic field propagation equations that satisfy the electromagnetic wave propagation equation of ground penetrating radar;

[0018] The Max equation used is as follows:

[0019]

[0020] In the formula, Indicates the curl of the magnetic field. ε0 represents the curl of the electric field, J represents the current density, and μ0 represents the permeability.

[0021] Step 4-2) Combine the decoupled Maxwell equations with the predicted electric field and dielectric constant to construct the physical residual formula that satisfies the electromagnetic wave propagation equation.

[0022] The constructed physical residual formula is as follows:

[0023]

[0024] In the formula, E is the electric field, σ is the conductivity, ε is the permittivity, μ is the permeability, and S is the magnetic field. E It is an electric field source term;

[0025] Step 4-3) Integrate the constructed physical residual formula as the physical loss function into the MSCCEAUNet framework to achieve joint data and physical-driven inversion, ensuring that the total loss of the MSCCEAUNet model is minimized.

[0026] In step five, the following steps are used to determine the optimal weights for the loss of each part of MSCCEAUNet:

[0027] Step 5-1) Use hyperparameter optimization to automatically select the optimal weights for each part of the data loss, physical loss, and structural similarity loss of MSCCEAUNet, ensuring that the loss of each part is minimized.

[0028] The formula for the constructed training loss function is as follows:

[0029] L total =αL data +βL physical +γL SSIM ;

[0030] In the formula, α refers to the weight proportion of data loss, β refers to the weight proportion of physical loss, and γ refers to the weight proportion of structural similarity loss; L data It is data loss, L physical It is the loss of the physical equations, L SSIM It is structural similarity loss;

[0031] Step 5-2) After determining the optimal weight ratio of each part of the loss, set the weight parameters of each part of MSCCEAUNet to be jointly driven by the best starting data and physics.

[0032] In step two, the constructed MSCECAUNet model is specifically as follows:

[0033] The MSCECAUNet network includes the first Transformer Encoder (MSC) layer to the fifth Transformer Encoder (MSC) layer, the first Transformer Encoder layer to the fifth Transformer Encoder layer, the first ECA layer to the fourth ECA layer, the first Transformer Decoder (MSC) layer to the fifth Transformer Decoder (MSC) layer, the first Transformer Decoder layer to the fifth Transformer Decoder layer, and an FPC module;

[0034] The GPR B-Scan profile is used as input feature to the first Transformer Encoder MSC layer. The output of the first Transformer Encoder MSC layer is connected to the input of the first Transformer Encoder. The output of the first Transformer Encoder is connected to the input of the first ECA layer and the input of the second Transformer Encoder MSC layer. The output of the first ECA layer and the output of the second Transformer Decoder are connected to the input of the first Transformer Decoder MSC layer. The output of the first Transformer Decoder MSC layer is connected to the input of the first Transformer Decoder. The output of the first Transformer Decoder is connected to the input of the FPC module. The output of the FPC module is a dielectric constant distribution map.

[0035] The output of the second Transformer Encoder MSC layer is connected to the input of the second Transformer Encoder layer. The output of the second Transformer Encoder layer is connected to the input of the second ECA layer and the input of the third Transformer Encoder MSC layer. The output of the second ECA layer and the output of the third Transformer Decoder layer are connected to the input of the second Transformer Decoder MSC layer. The output of the second Transformer Decoder MSC layer is connected to the input of the second Transformer Decoder layer.

[0036] The output of the third Transformer Encoder MSC layer is connected to the input of the third Transformer Encoder layer; the output of the third Transformer Encoder layer is connected to the input of the third ECA layer and the input of the fourth Transformer Encoder MSC layer; the output of the third ECA layer and the output of the fourth Transformer Decoder layer are connected to the input of the third Transformer Decoder MSC layer; and the output of the third Transformer Decoder MSC layer is connected to the input of the third Transformer Decoder layer.

[0037] The output of the fourth Transformer Encoder MSC layer is connected to the input of the fourth Transformer Encoder layer; the output of the fourth Transformer Encoder layer is connected to the input of the fourth ECA layer and the input of the fifth Transformer Encoder MSC layer; the output of the fourth ECA layer and the output of the fifth Transformer Decoder layer are connected to the input of the fourth Transformer Decoder MSC layer; and the output of the fourth Transformer Decoder MSC layer is connected to the input of the fourth Transformer Decoder layer.

[0038] The output of the fifth Transformer Encoder MSC layer is connected to the input of the fifth Transformer Encoder layer, the output of the fifth Transformer Encoder layer is connected to the input of the fifth Transformer Decoder MSC layer, and the output of the fifth Transformer Decoder MSC layer is connected to the input of the fifth Transformer Decoder layer.

[0039] The Transformer Encoder MSC layer of the MSCECAUNet model includes a single-kernel convolutional neural network layer SK-ConvNet, a three-dimensional convolutional neural network layer 3D-ConvNet, a two-dimensional convolutional neural network layer C2-ConvNet, a three-dimensional convolutional neural network layer C3-ConvNet, and an FFM module.

[0040] The input features are fed into the SK-ConvNet, 3D-ConvNet, C2-ConvNet, and C3-ConvNet layers, respectively. Then, the outputs of the above convolutional neural network layers are connected to the input of the FFM module, the output of the FFM module is connected to the input of the 3D-ConvNet, and the output of the 3D-ConvNet is connected to the input of the Kth Transformer Encoder layer of the MSCECOAUNet model; where K represents the number of layers in the MSCECOAUNet model.

[0041] The ECA layer of the MSCECAUNet model includes the input feature module IFM, the global average pooling module GAPM, the channel number adaptive K value module CKM, the K×K one-dimensional convolution module K×K 1D Conv, the sigmoid activation function module SAFM, and the output feature module OFM.

[0042] The output of the Transformer Encoder layer is connected to the input of the ECA layer and fed into the Input Feature Module (IFM). The output of the IFM is connected to the input of the Global Average Pooling Module (GAPM). The output of GAPM is connected to the input of the Channel Adaptive K-value Module (CKM). The output of CKM is connected to the input of the K×K 1D Conv Module (K×K 1D Conv). The output of K×K 1D Conv is connected to the input of the Sigmoid Activation Function Module (SAFM). The output of SAFM is connected to the input of the Output Feature Module (OFM). The output of OFM is connected to the input of the Kth Transformer Decoder MSC layer.

[0043] Compared with the prior art, the present invention has the following technical effects:

[0044] 1) The network of this invention adopts the traditional U-Net network as the framework and the transformer encoder-decoder as the basic structure. It integrates multi-scale cascaded convolution modules and efficient channel attention mechanism modules. This network can effectively extract the key features of termite nests in earth-rock dams and achieve accurate location of termite nests.

[0045] 2) This invention introduces physical information constraints: Maxwell's equations, which conform to the propagation of electromagnetic waves underground, are added to the deep learning model as physical constraints, thereby ensuring the accuracy and physical interpretability of the termite nest inversion results of earth-rock dams.

[0046] 3) This invention is based on automated annotation of finite difference simulation data: Geological modeling is performed using gprMax software, automatically generating annotation data. This method sets the simulation size, spatial step size, and the location, size, and dielectric constant of subsurface anomalies. Using these known parameters, subsurface structure maps can be automatically annotated, significantly reducing the time and workload of manual annotation. This innovative technology provides accurate and efficient annotation data for termite nest detection in earth-rock dams, significantly improving the training effect and prediction accuracy of inversion models, and promoting the application and development of intelligent detection technology.

[0047] 4) The intelligent system for health monitoring of earth-rock dams of the present invention: This model can be integrated into the health monitoring system of earth-rock dams to realize real-time monitoring and early warning of termite nest expansion, and ensure the long-term safety of the dam structure. Attached Figure Description

[0048] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0049] Figure 1 This is an overall flowchart of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of the MSCCEAUNet model in this invention;

[0051] Figure 3 This is a schematic diagram of the efficient channel attention mechanism (ECA) module in this invention;

[0052] Figure 4 This is a schematic diagram of the structure of the multi-scale cascaded convolutional module (MSC) in this invention;

[0053] Figure 5 This is an actual geological model diagram in an embodiment of the present invention;

[0054] Figure 6 This is a B-Scan waveform diagram corresponding to the geological model in this embodiment of the invention;

[0055] Figure 7 This is a schematic diagram of the underground dielectric constant inversion structure in an embodiment of the present invention;

[0056] Figure 8 This is a schematic diagram illustrating the training loss of a deep learning model without physical loss in this invention. Detailed Implementation

[0057] A termite nest detection method for earth-rock dams, jointly driven by MSCCEAUNet and physical constraints, includes the following steps:

[0058] Step 1: Simulate a complex geological model of an earth-rock dam structure with termite nests, and use the finite difference method (FDTD) to generate corresponding b-scan data and underground dielectric constant data as simulation datasets;

[0059] Step 2: Construct an MSCCEAUNet network with multi-scale concatenated convolution (MSC) and efficient channel attention mechanism (ECA), and input the complex geological model data generated by the simulation in Step 1 into the MSCCEAUNet network for training;

[0060] Step 3: Add Gaussian noise and random medium perturbation data with different signal-to-noise ratios to train MSCCEAUNet again. Based on the training results, optimize the MSCCEAUNet structure to improve the model's anti-interference ability and robustness.

[0061] Step 4: Integrate the physical residual equations that satisfy the physical information constraints of Maxwell's equations for ground-penetrating radar into the MSCCEAUNet framework as physical loss functions to achieve joint data-driven inversion.

[0062] Step 5: Add data with different signal-to-noise ratios of Gaussian noise and random medium perturbations again to perform joint data and physics-driven training, and determine the optimal weights for each part of the data loss, physical loss and structural similarity loss of MSCCEAUNet.

[0063] In step one, the gprMax simulation software is used to simulate the earth-rock dam structure model with termite nests.

[0064] In step four, when jointly driving the model and physical information constraints, the following steps are adopted:

[0065] Step 4-1) First, decouple the electric field and magnetic field propagation equations that satisfy the electromagnetic wave propagation equation of ground penetrating radar;

[0066] The Max equation used is as follows:

[0067]

[0068] In the formula, Indicates the curl of the magnetic field. ε0 represents the curl of the electric field, J represents the current density, and μ0 represents the permeability.

[0069] Step 4-2) Combine the decoupled Maxwell equations with the predicted electric field and dielectric constant to construct the physical residual formula that satisfies the electromagnetic wave propagation equation.

[0070] The constructed physical residual formula is as follows:

[0071]

[0072] In the formula, E is the electric field, σ is the conductivity, ε is the permittivity, μ is the permeability, and S is the magnetic field. E It is an electric field source term;

[0073] Step 4-3) Integrate the constructed physical residual formula as the physical loss function into the MSCCEAUNet framework to achieve joint data and physical-driven inversion, ensuring that the total loss of the MSCCEAUNet model is minimized.

[0074] In step five, the following steps are used to determine the optimal weights for the loss of each part of MSCCEAUNet:

[0075] Step 5-1) Use hyperparameter optimization to automatically select the optimal weights for each part of the data loss, physical loss, and structural similarity loss of MSCCEAUNet, ensuring that the loss of each part is minimized.

[0076] The formula for the constructed training loss function is as follows:

[0077] L total =αL data +βL physical +γL SSIM ;

[0078] In the formula, α refers to the weight proportion of data loss, β refers to the weight proportion of physical loss, and γ refers to the weight proportion of structural similarity loss; L data It is data loss, L physical It is the loss of the physical equations, L SSIM It is structural similarity loss;

[0079] Step 5-2) After determining the optimal weight ratio of each part of the loss, set the weight parameters of each part of MSCCEAUNet to be jointly driven by the best starting data and physics.

[0080] In step two, the constructed MSCECAUNet model is specifically as follows:

[0081] The MSCECAUNet network includes the first Transformer Encoder (MSC) layer to the fifth Transformer Encoder (MSC) layer, the first Transformer Encoder layer to the fifth Transformer Encoder layer, the first ECA layer to the fourth ECA layer, the first Transformer Decoder (MSC) layer to the fifth Transformer Decoder (MSC) layer, the first Transformer Decoder layer to the fifth Transformer Decoder layer, and an FPC module;

[0082] The GPR B-Scan profile is used as input feature to the first Transformer Encoder MSC layer. The output of the first Transformer Encoder MSC layer is connected to the input of the first Transformer Encoder. The output of the first Transformer Encoder is connected to the input of the first ECA layer and the input of the second Transformer Encoder MSC layer. The output of the first ECA layer and the output of the second Transformer Decoder are connected to the input of the first Transformer Decoder MSC layer. The output of the first Transformer Decoder MSC layer is connected to the input of the first Transformer Decoder. The output of the first Transformer Decoder is connected to the input of the FPC module. The output of the FPC module is a dielectric constant distribution map.

[0083] The output of the second Transformer Encoder MSC layer is connected to the input of the second Transformer Encoder layer. The output of the second Transformer Encoder layer is connected to the input of the second ECA layer and the input of the third Transformer Encoder MSC layer. The output of the second ECA layer and the output of the third Transformer Decoder layer are connected to the input of the second Transformer Decoder MSC layer. The output of the second Transformer Decoder MSC layer is connected to the input of the second Transformer Decoder layer.

[0084] The output of the third Transformer Encoder MSC layer is connected to the input of the third Transformer Encoder layer; the output of the third Transformer Encoder layer is connected to the input of the third ECA layer and the input of the fourth Transformer Encoder MSC layer; the output of the third ECA layer and the output of the fourth Transformer Decoder layer are connected to the input of the third Transformer Decoder MSC layer; and the output of the third Transformer Decoder MSC layer is connected to the input of the third Transformer Decoder layer.

[0085] The output of the fourth Transformer Encoder MSC layer is connected to the input of the fourth Transformer Encoder layer; the output of the fourth Transformer Encoder layer is connected to the input of the fourth ECA layer and the input of the fifth Transformer Encoder MSC layer; the output of the fourth ECA layer and the output of the fifth Transformer Decoder layer are connected to the input of the fourth Transformer Decoder MSC layer; and the output of the fourth Transformer Decoder MSC layer is connected to the input of the fourth Transformer Decoder layer.

[0086] The output of the fifth Transformer Encoder MSC layer is connected to the input of the fifth Transformer Encoder layer, the output of the fifth Transformer Encoder layer is connected to the input of the fifth Transformer Decoder MSC layer, and the output of the fifth Transformer Decoder MSC layer is connected to the input of the fifth Transformer Decoder layer.

[0087] The Transformer Encoder MSC layer of the MSCECAUNet model includes a single-kernel convolutional neural network layer SK-ConvNet, a three-dimensional convolutional neural network layer 3D-ConvNet, a two-dimensional convolutional neural network layer C2-ConvNet, a three-dimensional convolutional neural network layer C3-ConvNet, and an FFM module.

[0088] The input features are fed into the SK-ConvNet, 3D-ConvNet, C2-ConvNet, and C3-ConvNet layers, respectively. Then, the outputs of the above convolutional neural network layers are connected to the input of the FFM module, the output of the FFM module is connected to the input of the 3D-ConvNet, and the output of the 3D-ConvNet is connected to the input of the Kth Transformer Encoder layer of the MSCECOAUNet model; where K represents the number of layers in the MSCECOAUNet model.

[0089] The ECA layer of the MSCECAUNet model includes the input feature module IFM, the global average pooling module GAPM, the channel number adaptive K value module CKM, the K×K one-dimensional convolution module K×K 1D Conv, the sigmoid activation function module SAFM, and the output feature module OFM.

[0090] The output of the Transformer Encoder layer is connected to the input of the ECA layer and fed into the Input Feature Module (IFM). The output of the IFM is connected to the input of the Global Average Pooling Module (GAPM). The output of GAPM is connected to the input of the Channel Adaptive K-value Module (CKM). The output of CKM is connected to the input of the K×K 1D Conv Module (K×K 1D Conv). The output of K×K 1D Conv is connected to the input of the Sigmoid Activation Function Module (SAFM). The output of SAFM is connected to the input of the Output Feature Module (OFM). The output of OFM is connected to the input of the Kth Transformer Decoder MSC layer.

[0091] Example:

[0092] This embodiment includes the following steps:

[0093] Step 1: Use gprMax simulation software to simulate an earth-rock dam structure model with a termite nest, generating corresponding b-scan data and underground dielectric constant data as training data for the model. Below are the variable names and values ​​of each parameter given in the established complex geological model.

[0094] Table 1: Geological Model Parameters

[0095]

[0096]

[0097] Step 2: Input the complex geological model data of the simulated termite nest structure of the earth-rock dam into the MSCECAUNet model for training.

[0098] The MSCECAUNet model includes: an encoder-decoder network with a U-Net architecture, a multi-scale cascaded convolutional module (MSC), and an efficient channel attention mechanism module (ECA), which can effectively extract key features of termite nests in earth-rock dams.

[0099] (a) The encoder-decoder network in the U-Net architecture contains multi-scale cascaded convolutional modules (MSCs), with convolutional and deconvolution operations calculated using the following formulas:

[0100] 1. Input layer: The GPR B-scan signal is used as the input X0 of the model;

[0101] 2. Feature Fusion Layer: X0 undergoes a multi-resolution feature fusion operation after passing through a multi-scale cascaded convolutional module;

[0102] X1 = MSC(X0);

[0103] 3. Pooling layer: Performs max pooling on the feature map;

[0104] X2 = MaxPool(X1);

[0105] 4. Downsampling layer: Downsample each layer and apply the ReLU activation function.

[0106] Xn+1=Convn(Xn)+ReLU

[0107] After each downsampling and convolution operation, the number of channels changes as follows: 64->128->256->512->1024 channels.

[0108] 5. Bottleneck layer: In the deepest layer, convolution operations are applied to the 1024-channel feature map.

[0109] Xbottom = Conv(X1024) + ReLU

[0110] 6. Upsampling Convolution (Decoder Path)

[0111] The operations at each layer in the upsampling path are as follows:

[0112] 1) Use upconvolution (deconvolution) operations to increase spatial dimensions.

[0113] Xup = UpConv(Xbottom)

[0114] 2) Obtain feature maps from the downsampling path and perform feature fusion (skip connections).

[0115] Xconcat = Concat(Xeca, Xdown)

[0116] 3) Apply convolution operations and the ReLU activation function to reduce the number of channels.

[0117] Xconv = Conv(Xconcat) + ReLU

[0118] The number of channels is gradually halved during the upsampling process: 1024->512->256->128->64.

[0119] 7. Output layer: The output of this layer is used to predict the dielectric constant distribution.

[0120] Xfinal = Conv(X64)

[0121] (ii) Multi-scale Cascaded Convolutional Module (MSC)

[0122] This multi-scale cascaded convolutional module primarily performs multi-resolution feature fusion for a termite nest in an earth-rock dam. It mainly includes four types of convolutions with receptive field sizes of 1*1, 3*3, 5*5, and 7*7. Considering computational complexity, the 5*5 and 7*7 convolutions are replaced with two cascaded 3*3 convolutions and three cascaded 3*3 convolutions. Finally, all feature channel maps are fused together and then processed by a 3*3 convolutional module for overall feature map extraction.

[0123] (1) Input layer: Input B-scan signal data X0

[0124] (2) Branching operations:

[0125] X1 = Conv3 × 3(X0)

[0126] X2 = Conv3 × 3(X0)

[0127] X3=(Conv3×3)2(X0)

[0128] X4=(Conv3×3)3(X0)

[0129] Where X1, X2, X3, and X4 represent the features after convolution through four branches.

[0130] (3) Splicing layer:

[0131] Xconcat = Concat(X1, X2, X3, X4)

[0132] Where contact represents the concatenation operation.

[0133] (4) Final output layer:

[0134] Y = Conv3 × 3(Xconcat)

[0135] (III) The Efficient Channel Attention (ECA) module includes the following steps:

[0136] (1) Perform local cross-channel interaction between channel features and their nearest neighbor channel features. The calculation formula is as follows:

[0137]

[0138] Where |t odd represents the odd number closest to t; c represents the total number of channels; γ and b are fixed values, 2 and 1 respectively.

[0139] (2) Perform a weighted summation on all channels, using the following formula:

[0140]

[0141] In the formula: Representative and y i The set of k adjacent channels; To be with y i The j-th adjacent channel outputs; α j This represents the weight parameters shared by all channels; δ is the sigmoid function.

[0142] Step 3: Add Gaussian noise with different signal-to-noise ratios and random medium disturbance termite nest data to train MSCCEAUNet again. Based on the training results, optimize the MSCCEAUNet structure to improve the model's anti-interference ability and robustness.

[0143] Step 4: Integrate the physical residual equations that satisfy the physical information constraints of Maxwell's equations for ground penetrating radar into the MSCCEAUNet framework as a physical loss function to achieve joint data-driven inversion and ensure the inversion accuracy of termite nests in earth-rock dams.

[0144] Ground-penetrating radar (GPR) uses electromagnetic waves for detection. However, the propagation of these waves underground follows Maxwell's equations, which describe the interaction of electric and magnetic fields. These equations are fundamental to describing electromagnetic phenomena and can therefore guide data processing and output in deep learning models for Ground-Penetrating Radar (GPR) inversion. The propagation of electromagnetic waves is described by Maxwell's equations, and in GPR, the changes in the electric and magnetic fields are represented by the following equations:

[0145]

[0146]

[0147] In the formula Indicates the curl of the magnetic field. ε0 represents the curl of the electric field, J represents the current density, and μ0 represents the permeability.

[0148] The physical equations used in this method are constrained by the predicted electric field and dielectric constant, and the physical residuals of the electromagnetic wave propagation equations are introduced to ensure that the predicted results satisfy physical laws. The residual equations are shown below:

[0149]

[0150] In the formula, E is the electric field, σ is the conductivity, ε is the permittivity, μ is the permeability, and S is the magnetic field. E It is an electric field source term.

[0151] Step 5: Add termite nest data with varying signal-to-noise ratios and random medium disturbances again for joint data and physics-driven training, and determine the optimal weights for the data loss, physical loss, and structural similarity loss components of MSCCEAUNet. The total loss function formula is as follows:

[0152] L total =αL data +βL physical +γL SSIM

[0153] Among them, L data It is data loss, L physical It is the loss of the physical equations, L SSIM This is the structural similarity loss. α, β, and γ are the weight values ​​for the proportion of the three parts of the loss, respectively.

[0154] according to Figure 5 , Figure 6 , Figure 7 The actual underground model structure diagram of the termite nest in the earth-rock dam, along with the corresponding B-Scan waveform diagram and the underground dielectric constant structure diagram inverted using a deep learning model with added physical constraints, shows that the location, size, and shape of the inverted termite nest have a very high degree of similarity to the actual structure diagram. Furthermore, based on... Figure 8 The training loss diagram shows that adding physical constraints not only improves the effect and accuracy of deep learning model inversion, but also makes the underground dielectric constant value of the termite nest in the earth-rock dam very similar to the actual value, satisfying the physical interpretability of electromagnetic wave underground propagation. In summary, the ground-penetrating radar inversion method driven by model and physical information constraints demonstrates excellent performance, adapting to complex geological structures and effectively dealing with underground noise interference. This method provides an innovative technical path for the detection of termite nests in earth-rock dams, further improving the accuracy and reliability of dam monitoring and providing a solid guarantee for the long-term safety and stable operation of dams.

[0155] This invention addresses two key issues. First, it tackles the complexity of underground models and potential noise issues by establishing a complex geological model of termite nests in earth-rock dams and using data with added noise and random media disturbances. This results in a robust and widely applicable deep learning inversion model that effectively extracts waveform features of termite nests in earth-rock dams, ensuring accurate detection and location. Second, by appropriately incorporating physical information constraints into the model, it effectively improves the physical consistency of the inversion results, ensuring both inversion accuracy and physical interpretability in the detection of termite nests in earth-rock dams.

[0156] This invention employs the U-Net framework in its constructed deep learning model and introduces a multi-scale concatenated convolutional module (MSC) and an efficient channel attention mechanism (ECA), which not only improves the model's inversion performance but also enhances its robustness against interference. The method of introducing physical information constraints in this invention utilizes prediction results based on electric field and dielectric constant, incorporating the physical residuals of the electromagnetic wave propagation equation to ensure that the prediction results satisfy physical laws. This invention provides an innovative solution for the application of GPR in the field of ant nest detection in earth-rock dams. This method not only improves the accuracy of ant nest detection but also empowers intelligent monitoring and risk early warning of dam structures, providing technical assurance for the long-term safety and stable operation of dams.

Claims

1. A method for detecting termite nests in earth-rock dams based on MSCCEAUNet and physical constraints, characterized in that, Includes the following steps: Step 1: Simulate a complex geological model of an earth-rock dam structure with termite nests, and use the finite difference method (FDTD) to generate corresponding b-scan data and underground dielectric constant data as simulation datasets; Step 2: Construct the MSCCEAUNet network with multi-scale cascaded convolutional MSC and efficient channel attention mechanism ECA, and input the complex geological model data generated by the simulation in Step 1 into the MSCCEAUNet network for training; Step 3: Add Gaussian noise and random medium perturbation data with different signal-to-noise ratios to train MSCCEAUNet again. Based on the training results, optimize the MSCCEAUNet structure to improve the model's anti-interference ability and robustness. Step 4: Integrate the physical residual equations that satisfy the physical information constraints of Maxwell's equations for ground-penetrating radar into the MSCCEAUNet framework as physical loss functions to achieve joint data-driven inversion. Step 5: Add data with different signal-to-noise ratios of Gaussian noise and random medium perturbation again to perform joint data and physics-driven training, and determine the optimal weights of each part of MSCCEAUNet's data loss, physical loss and structural similarity loss; In step four, when jointly driving the model and physical information constraints, the following steps are adopted: Step 4-1) First, decouple the electric field and magnetic field propagation equations that satisfy the electromagnetic wave propagation equation of ground penetrating radar; The Max equation used is as follows: ； ； In the formula, Indicates the curl of the magnetic field. The curl of the electric field. Represents the dielectric constant. Indicates current density, Indicates magnetic permeability; Step 4-2) Combine the decoupled Maxwell equations with the predicted electric field and dielectric constant to construct the physical residual formula that satisfies the electromagnetic wave propagation equation. The constructed physical residual formula is as follows: ； In the formula, E is the electric field. It is conductivity. It is the dielectric constant. It is the permeability. It is an electric field source term; Step 4-3) Integrate the constructed physical residual formula as the physical loss function into the MSCCEAUNet framework to achieve joint data and physical-driven inversion, ensuring that the total loss of the MSCCEAUNet model is minimized.

2. The method according to claim 1, characterized in that, In step one, the gprMax simulation software is used to simulate the earth-rock dam structure model with termite nests.

3. The method according to claim 1, characterized in that, In step five, the following steps are used to determine the optimal weights for the loss of each part of MSCCEAUNet: Step 5-1) Use hyperparameter optimization to automatically select the optimal weights for each part of the data loss, physical loss, and structural similarity loss of MSCCEAUNet, ensuring that the loss of each part is minimized. The formula for the constructed training loss function is as follows: ； In the formula, This refers to the weighting percentage of data loss. This refers to the weighting percentage of physical losses. This refers to the weighting of structural similarity loss; L data It's data loss. It is a loss of the physical equations. It is structural similarity loss; Step 5-2) After determining the optimal weight ratio of each part of the loss, set the weight parameters of each part of MSCCEAUNet to be driven by the best starting data and physical parameters.

4. The method according to claim 1, characterized in that, In step two, the constructed MSCECAUNet model is specifically as follows: The MSCECAUNet network includes the first Transformer Encoder (MSC) layer to the fifth Transformer Encoder (MSC) layer, the first Transformer Encoder layer to the fifth Transformer Encoder layer, the first ECA layer to the fourth ECA layer, the first Transformer Decoder (MSC) layer to the fifth Transformer Decoder (MSC) layer, the first Transformer Decoder layer to the fifth Transformer Decoder layer, and an FPC module; The GPR B-Scan profile is used as input feature to the first Transformer Encoder MSC layer. The output of the first Transformer Encoder MSC layer is connected to the input of the first Transformer Encoder. The output of the first Transformer Encoder is connected to the input of the first ECA layer and the input of the second Transformer Encoder MSC layer. The output of the first ECA layer and the output of the second Transformer Decoder are connected to the input of the first Transformer Decoder MSC layer. The output of the first Transformer Decoder MSC layer is connected to the input of the first Transformer Decoder. The output of the first Transformer Decoder is connected to the input of the FPC module. The output of the FPC module is a dielectric constant distribution map. The output of the second Transformer Encoder MSC layer is connected to the input of the second Transformer Encoder layer. The output of the second Transformer Encoder layer is connected to the input of the second ECA layer and the input of the third Transformer Encoder MSC layer. The output of the second ECA layer and the output of the third Transformer Decoder layer are connected to the input of the second Transformer Decoder MSC layer. The output of the second Transformer Decoder MSC layer is connected to the input of the second Transformer Decoder layer. The output of the third Transformer Encoder MSC layer is connected to the input of the third Transformer Encoder layer; the output of the third Transformer Encoder layer is connected to the input of the third ECA layer and the input of the fourth Transformer Encoder MSC layer; the output of the third ECA layer and the output of the fourth Transformer Decoder layer are connected to the input of the third Transformer Decoder MSC layer; the output of the third Transformer Decoder MSC layer is connected to the input of the third Transformer Decoder layer. The output of the fourth Transformer Encoder MSC layer is connected to the input of the fourth Transformer Encoder layer; the output of the fourth Transformer Encoder layer is connected to the input of the fourth ECA layer and the input of the fifth Transformer Encoder MSC layer; the output of the fourth ECA layer and the output of the fifth Transformer Decoder layer are connected to the input of the fourth Transformer Decoder MSC layer; and the output of the fourth Transformer Decoder MSC layer is connected to the input of the fourth Transformer Decoder layer. The output of the fifth Transformer Encoder MSC layer is connected to the input of the fifth Transformer Encoder layer, the output of the fifth Transformer Encoder layer is connected to the input of the fifth Transformer Decoder MSC layer, and the output of the fifth Transformer Decoder MSC layer is connected to the input of the fifth Transformer Decoder layer.

5. The method according to claim 4, characterized in that, The TransformerEncoder MSC layer of the MSCECAUNet model includes a single-kernel convolutional neural network layer SK-ConvNet, a three-dimensional convolutional neural network layer 3D-ConvNet, a two-dimensional convolutional neural network layer C2-ConvNet, a three-dimensional convolutional neural network layer C3-ConvNet, and an FFM module. The input features are fed into the SK-ConvNet, 3D-ConvNet, C2-ConvNet, and C3-ConvNet layers, respectively. Then, the outputs of the above convolutional neural network layers are connected to the input of the FFM module, the output of the FFM module is connected to the input of the 3D-ConvNet, and the output of the 3D-ConvNet is connected to the input of the Kth Transformer Encoder layer of the MSCECOAUNet model; where K represents the number of layers in the MSCECOAUNet model.

6. The method according to claim 4, characterized in that, The ECA layer of the MSCECAUNet model includes the input feature module IFM, the global average pooling module GAPM, the channel number adaptive K value module CKM, the K×K one-dimensional convolution module K×K 1DConv, the sigmoid activation function module SAFM, and the output feature module OFM. The output of the Transformer Encoder layer is connected to the input of the ECA layer and fed into the Input Feature Module (IFM). The output of the IFM is connected to the input of the Global Average Pooling Module (GAPM). The output of GAPM is connected to the input of the Channel Adaptive K-value Module (CKM). The output of CKM is connected to the input of the K×K 1D Conv Module (K×K 1D Conv). The output of K×K 1D Conv is connected to the input of the Sigmoid Activation Function Module (SAFM). The output of SAFM is connected to the input of the Output Feature Module (OFM). The output of OFM is connected to the input of the Kth Transformer Decoder MSC layer.

Images